When he first started working at Bank of New York Mellon, Massimo Young, head of data solutions in investment management, had a philosophy of "machine learning is cool and I want to go apply it."
He made that very pitch to a senior executive, who threw it right back at him.
"That conversation didn't go very well," Young said in a panel discussion at the In|Vest conference Tuesday.
"The question is, why should we do that?" Young said. "We should start with where the problems are we want to solve. I had a number of false starts where we had use cases that turned out not to require machine learning at all."
In one example, the wealth management division wanted to monitor how far advisers were skewing away from the model portfolios they were given.
"There are good reasons to skew away from models, and there are bad reasons," Young said. "We wanted to understand risks being taken, where dispersion is from the model portfolios to the actual portfolios wealth managers were implementing. We thought this was a great use for clustering analysis. It ended up being much simpler than that from an analytical perspective."
Young's group went through a demanding data integration project to provide portfolio analytics, pulling data together from different source systems, understanding it, cleaning it, documenting it and making sure the business and information technology experts saw eye to eye on everything.
"That was the challenging part," he said. In the end, the new portfolio analytics were helpful — they just didn't require machine learning.
Andrew Brzezinski, vice president of Fidelity Institutional, told of similar experiences. He runs a data science team for the Fidelity Investments unit.
"In many instances the thing that is posed as an AI use case could be solved simply," he said.
His group tries to help financial advisers understand which clients they should be speaking to, and about what.
"Even simple analytics and data aggregations will already start providing good reference points for people to start answering those kinds of questions," Brzezinski said.
Later, machine learning use cases crop up: What is next thing I should focus on, who should I speak to next? What's the driver of this? Is there a life event of a client I should think about? How can we anticipate it and collect the data that will tell us about that? Is there risk of attrition? Can we anticipate that and give a score to it so they can have the right conversations?
Even with simpler data projects, his team always needs to lead the conversation with, where is the value? Brzezinski said.
"The question comes back, why? This is superexpensive; why are we doing this?" he said. "So leading with business value is great."
